Simulated annealing algorithm : technical improvements

Abstract

We present an overview of the main problem-independent sequential and parallel amelioration techniques of the Simulated Annealing algorithm. We begin by briefly exposing the theoretical framework encompassing the standard markovian model, the notion of cycle and the optimal temperature schedules. Theory of cycles yields explicit relationships between the geometry of the energy landscape and the expected behavior of the algorithm. It leads to the design of efficient temperature schedules, as well as to improvements of the algorithm behavior by distorting the cost function. Next, we present a survey of parallelization techniques, focussing on problem-independent synchronous strategies. They provide flexible and general tools, allowing operational research practitioners to take advantage of the computational power of parallel architectures. We conclude with an application. It concerns the search for Hamiltonian paths in cubic graphs. It brings to the fore the efficiency of the cost function distortions technique, when used in combination with Parallel Simulated Annealing.